Australian trials for renewables forecasting barking up wrong tree?

We wonder if the Australian Renewable Energy Agency has just gone and wasted a ton of money chasing weather forecasts? It has backed 11 different approaches to forecasting outputs for solar and wind energy projects, costing it AU$9 million.

We first noticed this when Vestas put out a release on its subsidiary Utopus to pilot a better forecast for Wind in partnership with Infigen Energy in 2 wind farms in South Australia. Turns out that this is a case of throwing multiple pieces of money at multiple approaches and finding out which ones are best.

The idea is laudable because getting intermittent energy resources like wind and solar to behave predictably, solves part of the problem of using these resources as part of a baseload. That means being able to turn off more coal and gas fired turbines – increasingly more expensive and of course dirtier.

The other major approach to achieving this is to store some of that intermittent energy using some form of grid scale battery (see other stories covered this week) and a Vestas spokesman told us that both of these two approaches are worth developing side by side.

But of course that implies extra battery costs for renewables or for grid players, and given that we are not long into an era where renewables have begun to not need subsidy against conventional fossil fuels, in many cases there is little room in the budget for extra battery resources. We suspect that as we see further improvements in solar, such as the widespread uptake of BiFacial panels, we will see a further improvement in parity costings, and many more installations will come with batteries attached, to smooth delivery of intermittent energy sources.

One of the biggest problems is not just working with weather forecasts, but predicting highly local weather, which requires another level of detail.

Each of the 11 studies have a slightly different approach; some with machine learning and local sensors and the local weather forecast; others with cloud cameras and satellite data; others using numerical weather prediction and mesoscale models; but none of the approaches use anything really new.

Recently we came across a company a few weeks back called ClimaCel, a US Boston based company which claims to take an IoT approach to flood prediction. It has totally re-invented the manner if which it collects data for weather forecasting. It has purchased access to microwave backhaul data for cellular base stations, and analyzed it for things like “rain fade” in the strength of the radio signal, using AI, and done the same for satellite signals. On top of this it has taken data from barometric sensors in phones, which are ostensibly there to improve GPS location, but which can also give out atmospheric pressure readings. Another factor is that when weather is cold, phones spend their batteries more quickly, so it has data on this too coming straight out of the phones.

It has access to car telematics data too, and can see which cars turn on their headlamps, windscreen wipers or fog lamps, all of which give really good indicators to what the weather is currently like now at that location. All of the data is real time, and delivered only in summary, so it has low value and does not identify anyone’s behavior. The system even uses road cameras which can literally “see” rainfall and the angle it is falling at, showing prevalent winds. All this is pushed through a big AI, which then takes all this highly “local” data, and tries to join it up for a large scale weather map – so if rain is being experienced up the road and the prevailing wind is sending it to the next solar farm, then it should be there shortly, same for cloud cover or fog.

Initially this approach has been built into a flood warning system, but it could easily be adapted to include sensors in solar and wind farms, a history of what the power outputs were when these conditions were prevalent, and localized weather could be highly predictable.

The problem with using AI with just local sensors at the farms, plus weather forecasting, is that there simply is not enough local data to be “certain” of any given forecast. Whereas ClimaCel claims to have access to 500 million sensors at a cost of almost nothing, doing it this way.

A spokesperson for Vestas told us that “The local energy operator can punish renewable suppliers if they are out in their forecasts, and they have to forecast 5 minutes ahead of time in Australia, whereas most countries have a 10 minute out prediction. Our own Utopus solution currently works on a 10 minute forecast of energy, and needs to be updated to within 5 minutes to work better in Australia.

What happens when the 11 pilots are all partially successful? We presume the best solutions will be in great demand in Australia, or will anyone know which ones are best? They may all do roughly as well as each other.

Solar panels are intermittent primarily because of cloud cover, shadowing from nearby shrubs at certain times of the day, by occluding one another on tracking systems, and of dirt collecting on the panel surface, or when rain and fog reduces visibility. For Wind, it is just a matter of the wind not blowing, or blowing from inconsistent directions, which reduces the output.

We know of no other authority which has gone to such lengths to predict the effect weather has on renewables, but it might have been a good idea to put ClimaCel in touch with one or two of them, because we found its ideas fairly revolutionary, especially since many local weather forecasts are known to take too long to input data for a full macro weather model in the 5 minute timeframe.

The 11 projects will account for around 45% of registered wind and solar capacity, a total of around 3.5 GW. It is hoped that the owners of these renewable projects will be better able to trade spare energy in the merchant market once they have their 5 minutes ahead forecast carried out more accurately.

The companies involved in the trials include Windlab, Industrial Monitoring & Control, Meridian Energy Australia, Solar and Storage Modelling, Advisian, DNV GL, Fulcrum 3D Pty, Vestas Australian Wind, Aeolius Wind Systems and Proa Analytics.